EU Issues First-Ever GMP Annex on Artificial Intelligence: 10 Key Compliance Requirements for Pharmaceutical AI

Brussels, July 7, 2025 — The European Commission and PIC/S (Pharmaceutical Inspection Co-operation Scheme) have jointly released the draft “GMP Annex 22: Artificial Intelligence (AI)”, launching a three-month public consultation (July 7–October 7, 2025).

As the world’s first GMP guideline dedicated to AI applications in the pharmaceutical industry, Annex 22 establishes a comprehensive compliance framework that brings algorithmic innovation under auditable, life-cycle regulation.

Source: European Commission Directorate-General for Health and Food Safety (DG SANTE) and Pharmaceutical Inspection Co-operation Scheme (PIC/S) — Draft “Annex 22: Artificial Intelligence” (Public Consultation, July 2025).

Photo by Dusan_Cvetanovic on Pixabay

A Regulatory Milestone: From “Innovation” to “Compliance”

Annex 22 defines a strict boundary for AI use in pharmaceutical manufacturing — applicable only to systems that directly affect patient safety, product quality, or data integrity.

Built on the concept of “static models with deterministic outputs”, the guideline excludes adaptive, probabilistic, or generative AI models in GMP-critical operations. However, limited use of human-supervised generative AI (HITL) is permitted for non-critical functions such as energy management or inventory forecasting.

Three Core Principles Underpin the Framework

All AI systems operating under GMP must follow three principles:

Cross-functional collaboration among SMEs, QA, data scientists, and IT specialists with qualified responsibilities.

Documentation traceability, ensuring that every stage — from training to deployment — is fully auditable.

Quality risk management, evaluating AI’s potential impact on patient safety and product quality.

Ten Compliance Requirements for Pharmaceutical AI

From model design to post-market monitoring, Annex 22 outlines ten mandatory checkpoints, including:

Intended Use Documentation before development;

Performance verification that must meet or exceed human baselines;

Independent, representative test data with complete traceability;

Explainable AI mechanisms (e.g., SHAP, LIME) for transparent decisions;

Confidence thresholds and human review triggers for uncertain predictions;

Continuous operational monitoring to detect model drift or configuration changes.

Photo by Tara Winstead on Pexels

Hillgene’s Perspective: Building Trustworthy AI for Cell & Gene Manufacturing

As an early adopter of GMP digitalization, Hillgene welcomes the release of Annex 22 as a long-awaited step toward standardizing AI-driven quality systems.

The company’s AI-assisted analytical and manufacturing platforms have already integrated explainability, data segregation, and model drift monitoring — aligning closely with the Annex’s expectations.

“AI can only accelerate cell and gene therapy manufacturing if it’s transparent and accountable,” said a Hillgene R&D spokesperson. “Annex 22 provides the regulatory clarity that allows innovation and compliance to progress together.”

Hillgene continues to invest in interpretable AI for critical quality attribute (CQA) prediction, digital batch record analysis, and risk-based process optimization, aiming to demonstrate how compliant AI can enhance both efficiency and regulatory confidence.

Photo by Navy Medicine on Unsplash

Compliance Roadmap and Industry Impact

Annex 22 recommends a phased implementation approach — from gap assessment and data restructuring to model validation and continuous monitoring.

Experts believe the guideline marks the beginning of a “hard-law era” for pharmaceutical AI, ending years of fragmented practice.

With Annex 22, the path forward is clear: AI in pharma must be interpretable, verifiable, and fully GMP-compliant — a direction Hillgene has already embraced.